Self-adjusting Models for Semi-supervised Learning...
Transcript of Self-adjusting Models for Semi-supervised Learning...
Ferit Akovaa,b
in collaboration with Yuan Qib, Bartek Rajwac and Murat Dundara
a Computer & Information Science Department, Indiana University –
Purdue University, Indianapolis (IUPUI) b Computer Science Department, Purdue University, West Lafayette, IN
c Discovery Park, Purdue University, West Lafayette, IN
Overview Semi-supervised learning and the fixed model assumption
2
Gaussian assumption per class
labeled unlabeled
Overview
A new direction for Semi-supervised learning
utilizes unlabeled data to improve learning even when labeled data is partially-observed
uses self-adjusting generative models instead of fixed ones
discovers new classes and new components of existing classes
3
Outline
1. Learning in Non-exhaustive Settings
2. Motivating Problems
3. Overview of the Proposed Approach
4. Partially-observed Hierarchical Dirichlet Processes
5. Illustration and Experiments
6. Conclusion and Future Work
4
Non-exhaustive Setting Training dataset is unrepresentative if the list of classes is
incomplete, i.e., non-exhaustive
Future samples of unknown classes will be misclassified (into one of the existing classes) with a probability one
ill-defined classification problem!
blue: known
green & purple: unknown 5
What may lead to non-exhaustiveness?
Some classes may not be in existence
Classes may exist but may not be known
Classes may be known but samples are unobtainable
Exhaustive training data not realistic for many problems
6
Some Application Domains
Classification of documents by topics
research articles, web pages, news articles
Image annotation
Object categorization
Bio-detection
Hyperspectral image analysis
7
Biodetection
Food Pathogens Acquired samples are from most
prevalent classes
High mutation rate, new classes can emerge anytime
An exhaustive training library simply impractical
Inherently non-exhaustive setting
A B
C D
(A) Listeria monocytogenes 7644,
(B) E. coli ETEC O25,
(C)Staphylococcus aureus P103,
(D)Vibrio cholerae O1E
8
Hyperspectral Data Analysis Military projects, GIS, urban planning, ... Physically inaccessible or dynamically changing areas
Enemy territories, special military bases urban fields, construction areas
Impractical to obtain an exhaustive training data
9
Semi-supervised Learning (SSL) Traditional approaches
1. self-training, 2. co-training, 3. transductive methods, 4. graph-based methods, 5. generative mixture models
Unlabeled data improves classification under certain conditions, but primarily:
model assumption matches the model generating the data
Limited labeled data not only scarce, but usually data distribution not fully represented or maybe evolving
10
SSL in Non-exhaustive Settings
A new framework for semi-supervised learning
replaces the (brute-force fitting of a) fixed data model
dynamically includes new classes/components
classifies incoming samples more accurately
A self-adjusting model to better accommodate unlabeled data
11
Our Approach in a Nutshell
Classes as Gaussian mixture model (GMM) with unknown number of components
Extension of HDP to dynamically model new components/classes
Parameter sharing across inter- & intra-class components
Collapsed Gibbs sampler for inference
12
Our Notation
13
DP, HDP Briefly… Dirichlet Process (DP): a nonparametric prior over the number of
mixture components with base distribution G0 and parameter α
Hierarchical DP: models each group/class as a DP mixture and couples the Gj’s through a higher level DP
𝑥𝑗𝑖|𝜃𝑗𝑖 ~ 𝑝(⋅ |𝜃𝑗𝑖) 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝑗, 𝑖
𝜃𝑗𝑖|𝐺𝑗 ~ 𝐺𝑗 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝑗, 𝑖
𝐺𝑗|𝐺0, 𝛼 ~ 𝐷𝑃(𝐺0, 𝛼) 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝑗
𝐺0|𝐻, 𝛾 ~ 𝐷𝑃(𝐻, 𝛾)
α controls the prior probability of a new component
14
Modeling with HDP Chinese Restaurant Franchise (CRF) analogy
Restaurants correspond to classes, tables to mixture components and dishes in the “global menu” to unique parameters
First customer at a table orders a dish
Popular dishes more likely to be chosen
Role of γ in picking a new dish from the menu
15
Conditional Priors in CRF
Seating customers and assigning dishes to tables
tji – index of the table for customer i in restaurant j
kjt – index of the dish served at table t in restaurant j
𝑡𝑗𝑖|𝑡𝑗1, … , 𝑡𝑗,𝑖−1, 𝛼 ~𝛼
𝑛𝑗 + 𝛼𝛿𝑡𝑛𝑒𝑤 +
𝑚𝑗.
𝑡=1
𝑛𝑗𝑡𝑛𝑗 + 𝛼
𝛿𝑡
𝑘𝑗𝑡|𝑘𝑗1, … , 𝑘𝑗,𝑡−1, 𝛾 ~𝛾
𝑚.. + 𝛾𝛿𝑘𝑛𝑒𝑤 +
𝐾
𝑘=1
𝑚.𝑘𝑚.. + 𝛾
𝛿𝑘
16
Inference in HDP Gibbs sampler to iteratively sample the indicator variables
for tables and dishes given the state of all others
𝐭 = 𝑡𝑗𝑖 𝑖=1𝑛𝑗
𝑗=1
𝐽, 𝐤 = 𝑘𝑗𝑡 𝑡=1
𝑚𝑗.
𝑗=1
𝐽, 𝜙 = 𝜙𝑘 𝑘=1
𝐾
Conjugate pair of H and P(.|φ) allows for integrating out φ
to obtain a collapsed version
α and γ also sampled in each sweep based on number of
tables and dishes, respectively. (Escobar & West, 1994)
17
Gibbs Sampler for t and k Conditional weighted by number of samples
Joint probability weighted by number of components
18
Defining Partially-observed Setting
Observed classes/subclasses: Those initially available in the training library.
Unobserved classes/subclasses: Those not represented in the training library
New classes: classes discovered online, verified offline
limited to a single component until manual verification
19
HDP in a Partially-observed Setting
Two tasks:
1. Inferring component membership of labeled samples
2. Inferring both the group and component membership of unlabeled samples
Unlabeled samples evaluated for all existing components
20
Inference in Partially-observed HDP Updated Gibbs sampling inference for tji
21
Inference in Partially-observed HDP Updated inference for kjt for existing and new classes
22
Gaussian Mixture Model Data
Σ0, 𝑚, 𝜇0, 𝜅 estimated from labeled data by Empirical Bayes
23
Inference from GMM Data
24
Parameter Sharing in a GMM
25
Illustrative Example 3 classes as a mixture of 3 components
110 samples in each component, 10 randomly selected as labeled 100 considered as unlabeled
Covariance matrices from a set of 5 templates
26
Illustrative Example Standard HDP using only labeled data
A fixed generative model assigning full weight to labeled samples and reduced weight to unlabeled ones.
SA-SSL using labeled and unlabeled data with parameter sharing 27
1
2
3
Experiments - Evaluated Classifiers
Baseline supervised learning methods using only labeled data
Naïve-Bayes (SL-NB), Maximum likelihood (SL-ML), expectation maximization (SL-EM)
Benchmark semi-supervised learning methods
Self-training with base learners ML and NB (SELF)
Co-training with base learners ML and NB (CO-TR)
SSL-EM: Standard generative model approach
SSL-MOD: EM based approach with unobserved class modeling
SA-SSL: Proposed Self-adjusting SSL approach
28
Experiments – Classifier Design
Split available labeled data into train, unlabeled and test
Stratified sampling to represent each class proportionally
Consider some classes “unobserved” moving their samples from training set to unlabeled set
Non-exhaustive training set, exhaustive unlabeled and test sets
29
Experiments – Evaluation
Overall classification accuracy
Average accuracies on observed and unobserved classes
Newly created components associated with unobserved classes according to majority of samples
Repeated with 10 random test/train/unlabeled splits
30
Remote Sensing
31
20 components and 10 unique covariance matrices in total
Two to three components per each of the 8 classes
Half of the components shares covariance matrices
Remote Sensing Results
32
Pathogen Detection Experiment Total of 2054 samples from 28 bacteria classes
Each class contains between 40 to 100 samples
22 feature samples
4 classes made unobserved, 24 classes remains observed
30% as test, 20% as train and remaining 50% as unlabeled
Totally 180 components, 150
unique covariance matrices
Five to six components
per each class
One sixth of the components
shared parameter with others
33
Method Acc Acc-O Acc-U
SA-SSL 0.81 0.80 0.84
SSL-EM 0.64 0.75 0
SSL-MOD 0.67 0.74 0.26
SELF 0.59 0.70 0
CO-TR 0.60 0.72 0
SL-ML 0.62 0.73 0
SL-NB 0.52 0.62 0
SL-EM 0.30 0.35 0
Recap of the Contributions A new approach to learning with a non-exhaustively
defined labeled data set
A unique framework to utilize unlabeled samples in partially-observed semi-supervised settings
1) Extension of HDP model to entertain unlabeled data and to discover & recover new classes
2) Fully Bayesian treatment of mixture components to allow parameter sharing across different components
a) addresses the curse of dimensionality
b) connects observed classes with unobserved ones 34
Future Work
Replace Gibbs sampler with more scalable approximate inference methods
Speed-up for real-time analysis of sequential data via a sequential MCMC sampler
Extend the framework to hierarchically-structured datasets to associate discovered classes with higher level groups of classes
35